mlir::cast<quant::QuantizedType>(non_bias_type.getElementType());
      non_bias_types.push_back(non_bias_ele_type);
    } else {
      // The non-bias hasn't been quantized, let's skip this bias.
      break;
    }
  }
  // The non-bias hasn't been quantized, let's skip this bias.
  if (non_bias_types.size() != non_biases.size()) return {};

  return func(/*op_types=*/non_bias_types, /*adjusted_quant_dim=*/-1,

        auto attr = rewriter.getZeroAttr(type);
        bias = rewriter.create<arith::ConstantOp>(add_op.getLoc(), type, attr);
        auto none_af = rewriter.getStringAttr("NONE");

        bias =
            rewriter.create<AddOp>(add_op.getLoc(), bias, constant_val, none_af)
                .getOutput();
      } else {
        // If there no pre-existing bias and the `constant_val` is 1D, simply

def UndoBroadcastFullyConnectedBiasAdd : Pat<
  (TFL_AddOp $lhs, (Arith_ConstantOp:$const_value $bias), TFL_AF_None),
  (TFL_AddOp $lhs, (Arith_ConstantOp (FlattenTo1D $bias)), TFL_AF_None),
  [(AnyStaticShapeTensor $lhs),
   (IsLastDimEqualToNumElements $bias, $bias),
   (HasRankAtMost<4> $bias),
   (HasRankAtLeast<2> $bias),
   (IsDefinedByFullyConnectedOp $lhs),

  }
}

// Creates a new `tfl.qconst` op for the bias. The bias values are 0s, because
// this bias a dummy bias (note that bias fusion is not considered for this
// transformation). The quantization scale for the bias is input scale *
// filter scale. `filter_const_op` is used to retrieve the filter scales and
// the size of the bias constant.
TFL::QConstOp CreateTflConstOpForDummyBias(

    }
    const int64_t cost_per_col = 2 * weight_type.getNumElements();

    *count = cost_per_col * cols;

    auto bias = op->getOperand(2);
    if (bias) {
      auto bias_type =
          mlir::dyn_cast_or_null<mlir::RankedTensorType>(bias.getType());
      if (bias_type && bias_type.hasStaticShape()) {
        *count += output_type.getNumElements();
      }
    }

    return true;
  }

// CHECK: return %[[TRANSPOSE_1]]

// -----

// Tests that a `add(convolution(%activation, %weight), %bias)` pattern with the
// activation tensor of NCHW format and non-constant bias is converted to NHWC
// convolution, but without the deferred transpose for `stablehlo.add`.
// Transpose ops are inserted to the activation and output of

def FoldQuantWeightsIntoTposeConv : Pat<
  (TFL_TransposeConvOp
    $output_shape,
    (TFL_DequantizeOp $quant_weights),
    $quant_input,
    $bias, $padding, $stride_h, $stride_w, $faf),
  (TFL_TransposeConvOp $output_shape, $quant_weights,
    $quant_input, $bias, $padding, $stride_h, $stride_w, $faf), 

  %weight = arith.constant dense_resource<__elided__> : tensor<2x3x3x2xf32>
  %bias = arith.constant dense<[7.11401462, 7.05456924]> : tensor<2xf32>

  %q_input= "quantfork.qcast"(%input) : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3x!quant.uniform<i8:f32, 0.58810077742034317:-128>>

      bias_fn: Optional[ops.Operation] = None,
      activation_fn: Optional[ops.Operation] = None,
  ) -> module.Module:
    class MatmulModel(module.Module):
      """A simple model with a single matmul.

      Bias and activation function are optional.
      """

      def __init__(
          self,
          weight_shape: Sequence[int],
      ) -> None:
        """Initializes a MatmulModel.

        Args:

                         PatternRewriter& rewriter) {
  auto bias = op->getOperand(bias_idx);

  if (!mlir::isa<NoneType>(bias.getType())) return failure();

  // Proceed to create a zero bias.
  auto output = op->getResult(0);
  auto output_type = mlir::dyn_cast_or_null<RankedTensorType>(output.getType());
  if (!output_type) return failure();

  // bias should be a vector sized of the last output dim.